[mlir][tosa] Fix tosa-reduce-transposes to handle large constants better #148755
Conversation
|
Thank you for submitting a Pull Request (PR) to the LLVM Project! This PR will be automatically labeled and the relevant teams will be notified. If you wish to, you can add reviewers by using the "Reviewers" section on this page. If this is not working for you, it is probably because you do not have write permissions for the repository. In which case you can instead tag reviewers by name in a comment by using If you have received no comments on your PR for a week, you can request a review by "ping"ing the PR by adding a comment “Ping”. The common courtesy "ping" rate is once a week. Please remember that you are asking for valuable time from other developers. If you have further questions, they may be answered by the LLVM GitHub User Guide. You can also ask questions in a comment on this PR, on the LLVM Discord or on the forums. |
|
@llvm/pr-subscribers-mlir Author: Hanumanth (Hanumanth04) ChangesThe existing --tosa-reduce-transposes implementation is slow when transposing large constants (like weights in Linear operators) because it constructs expensive attributes during the transpose process. For example, with a 1000×401408 weight tensor (network screen shot at the end for reference) , the current implementation takes: <img width="1306" height="584" alt="image" src="https://github.com/user-attachments/assets/2ce7d7ca-7037-4ffa-9731-c7c756c32489" /> This change addresses the performance issue by working directly with the raw tensor data, eliminating the need for costly intermediate attributes. With the changes made in the PR for the same tensor mentioned above, we take: <img width="1328" height="568" alt="image" src="https://github.com/user-attachments/assets/5c639f02-8d9d-4572-b590-e82ed594d295" /> Network screen shot where this issue was observed: <img width="504" height="916" alt="image" src="https://github.com/user-attachments/assets/c16f5c3d-e1eb-4963-9fe3-de8f5127d3d2" /> Tagging here @sjarus and @sahas3 as I don't have the permission to add reviewers. Full diff: https://github.com/llvm/llvm-project/pull/148755.diff 1 Files Affected:
diff --git a/mlir/lib/Dialect/Tosa/Transforms/TosaReduceTransposes.cpp b/mlir/lib/Dialect/Tosa/Transforms/TosaReduceTransposes.cpp
index 8ebbbc94eb6a2..db7a3c671dedc 100644
--- a/mlir/lib/Dialect/Tosa/Transforms/TosaReduceTransposes.cpp
+++ b/mlir/lib/Dialect/Tosa/Transforms/TosaReduceTransposes.cpp
@@ -178,10 +178,8 @@ std::optional<DenseElementsAttr>
TosaReduceTransposes::transposeDenseAttribute(DenseElementsAttr input,
ArrayRef<int32_t> perms) {
RankedTensorType oldType = llvm::cast<RankedTensorType>(input.getType());
- RankedTensorType newType =
- RankedTensorType::get(applyTOSAPermutation(oldType.getShape(), perms),
- oldType.getElementType());
- size_t rank = oldType.getRank();
+ ArrayRef<int64_t> oldShape = oldType.getShape();
+ int64_t rank = oldType.getRank();
// Asserted by TransposeOp verifier and TOSA disallowing tensor with dimension
// 0. If not in place, something is very wrong.
@@ -190,65 +188,83 @@ TosaReduceTransposes::transposeDenseAttribute(DenseElementsAttr input,
return std::nullopt;
}
- if (input.isSplat())
+ auto newShape = applyTOSAPermutation(oldShape, perms);
+ RankedTensorType newType =
+ RankedTensorType::get(newShape, oldType.getElementType());
+
+ if (input.isSplat()) {
return input.reshape(newType);
+ }
+
+ auto rawData = input.getRawData();
+ if (!rawData.data()) {
+ return std::nullopt;
+ }
// The algorithm is approximately as follows:
- // input: perms, input flat array, input tensor type
- // (1/2) determine the strides of input/output if
- // they were strided in row-major order. (3) adjust the strides for the
- // input to be in the same order of indices as the output is written.
- // (4) process dimension by dimension. example: perms 2, 0, 1; input
- // 2x3x4; output 4x2x3 for i ... 4, j ... 2, k ... 3: output[i][j][k] =
- // input[j][k][i] output[6i + 3j + k] = input[12j + 4k + i] and we adjust
- // input strides to be as input[i + 12j + 4k] so we may process
- // layer-by-layer.
-
- // Step 1/2: Strides for input. We ignore output since row-major and can just
- // push_back.
-
- SmallVector<int64_t> originalInputStrides(rank);
- originalInputStrides[rank - 1] = 1;
- // index with int64_t to avoid overflow
- for (int64_t i = rank - 2; i >= 0; i--)
- originalInputStrides[i] =
- originalInputStrides[i + 1] * oldType.getDimSize(i + 1);
-
- // Step 3: Transpose strides of input to be same indexing (i, j, k, ...) as
- // output which is done in row-major order.
-
- SmallVector<int64_t> newInputStrides;
- newInputStrides.reserve(rank);
- for (int32_t v : perms)
- newInputStrides.push_back(originalInputStrides[v]);
-
- // Step 4: Write out the transposed "flat array" dimension by dimension.
-
- auto inputArray = input.getValues<Attribute>();
- SmallVector<std::pair<int64_t, int64_t>> boundsAndStrides;
- for (size_t i = 0; i < rank; i++)
- boundsAndStrides.push_back({newType.getDimSize(i), newInputStrides[i]});
-
- SmallVector<Attribute> resultArray;
- resultArray.reserve(inputArray.size());
-
- std::function<void(int64_t,
- SmallVector<std::pair<int64_t, int64_t>>::const_iterator)>
- processTransposeDim = [&](auto accumulatedIndex, auto it) {
- if (it == boundsAndStrides.end()) {
- resultArray.push_back(inputArray[accumulatedIndex]);
- return;
- }
-
- for (int64_t i = 0; i < it->first; i++) {
- int64_t j = accumulatedIndex + i * it->second;
- processTransposeDim(j, it + 1);
- }
- };
-
- processTransposeDim(0, boundsAndStrides.begin());
-
- return DenseElementsAttr::get(newType, resultArray);
+ // 1. Determine the strides of both input and output tensors in row-major
+ // order
+ // 2. Iterate through the output tensor linearly.
+ // 3. For each output position, decompose the linear index into
+ // multi-dimensional coordinates using output strides.
+ // 4. Use the permutation to map output coordinates to input coordinates and
+ // calculate the source linear index.
+
+ // Example: perms [2, 0, 1]; input 2x3x4; output 4x2x3
+ // for output linear index 11: decompose to output[1][1][2]
+ // using output strides [6,3,1]. Map to input coordinates using
+ // perms: dim 0→2, dim 1→0, dim 2→1, giving source position
+ // calculated as 1*inputStrides[2] + 1*inputStrides[0] + 2*inputStrides[1]
+ // = 1*1 + 1*12 + 2*4 = 21
+
+ size_t elementSize = oldType.getElementTypeBitWidth() / 8;
+ int64_t numElements = oldType.getNumElements();
+
+ SmallVector<char> outputBuffer(numElements * elementSize);
+ const char *inputPtr = rawData.data();
+ char *outputPtr = outputBuffer.data();
+
+ auto calculateStrides = [](ArrayRef<int64_t> shape) -> SmallVector<int64_t> {
+ int64_t rank = shape.size();
+ SmallVector<int64_t> strides(rank);
+ strides[rank - 1] = 1;
+ for (int64_t i = rank - 2; i >= 0; --i) {
+ strides[i] = strides[i + 1] * shape[i + 1];
+ }
+ return strides;
+ };
+
+ // Calculate strides for both input and output tensors
+ SmallVector<int64_t> inputStrides = calculateStrides(oldShape);
+ SmallVector<int64_t> outputStrides = calculateStrides(newShape);
+
+ auto mapCoordinates = [&](int64_t destLinearIndex) -> int64_t {
+ int64_t tempDestIndex = destLinearIndex;
+ int64_t sourceLinearIndex = 0;
+
+ // Decompose linear destination index into multi-dimensional
+ // coordinates dividing by output strides.
+ // Simultaneously map these coordinates through the permutation
+ // to calculate the corresponding source linear index.
+ for (auto j : llvm::seq<int64_t>(rank)) {
+ int64_t destCoord = tempDestIndex / outputStrides[j];
+ tempDestIndex %= outputStrides[j];
+ sourceLinearIndex += destCoord * inputStrides[perms[j]];
+ }
+
+ return sourceLinearIndex;
+ };
+
+ for (auto destLinearIndex : llvm::seq<int64_t>(numElements)) {
+ int64_t sourceLinearIndex = mapCoordinates(destLinearIndex);
+
+ // Copy the element from source to destination using type-agnostic byte
+ // copying.
+ std::memcpy(outputPtr + destLinearIndex * elementSize,
+ inputPtr + sourceLinearIndex * elementSize, elementSize);
+ }
+
+ return DenseElementsAttr::getFromRawBuffer(newType, outputBuffer);
}
// The SetVector should only contain ConstOp, ReshapeOp, TransposeOp
|
|
@llvm/pr-subscribers-mlir-tosa Author: Hanumanth (Hanumanth04) ChangesThe existing --tosa-reduce-transposes implementation is slow when transposing large constants (like weights in Linear operators) because it constructs expensive attributes during the transpose process. For example, with a 1000×401408 weight tensor (network screen shot at the end for reference) , the current implementation takes: <img width="1306" height="584" alt="image" src="https://github.com/user-attachments/assets/2ce7d7ca-7037-4ffa-9731-c7c756c32489" /> This change addresses the performance issue by working directly with the raw tensor data, eliminating the need for costly intermediate attributes. With the changes made in the PR for the same tensor mentioned above, we take: <img width="1328" height="568" alt="image" src="https://github.com/user-attachments/assets/5c639f02-8d9d-4572-b590-e82ed594d295" /> Network screen shot where this issue was observed: <img width="504" height="916" alt="image" src="https://github.com/user-attachments/assets/c16f5c3d-e1eb-4963-9fe3-de8f5127d3d2" /> Tagging here @sjarus and @sahas3 as I don't have the permission to add reviewers. Full diff: https://github.com/llvm/llvm-project/pull/148755.diff 1 Files Affected:
diff --git a/mlir/lib/Dialect/Tosa/Transforms/TosaReduceTransposes.cpp b/mlir/lib/Dialect/Tosa/Transforms/TosaReduceTransposes.cpp
index 8ebbbc94eb6a2..db7a3c671dedc 100644
--- a/mlir/lib/Dialect/Tosa/Transforms/TosaReduceTransposes.cpp
+++ b/mlir/lib/Dialect/Tosa/Transforms/TosaReduceTransposes.cpp
@@ -178,10 +178,8 @@ std::optional<DenseElementsAttr>
TosaReduceTransposes::transposeDenseAttribute(DenseElementsAttr input,
ArrayRef<int32_t> perms) {
RankedTensorType oldType = llvm::cast<RankedTensorType>(input.getType());
- RankedTensorType newType =
- RankedTensorType::get(applyTOSAPermutation(oldType.getShape(), perms),
- oldType.getElementType());
- size_t rank = oldType.getRank();
+ ArrayRef<int64_t> oldShape = oldType.getShape();
+ int64_t rank = oldType.getRank();
// Asserted by TransposeOp verifier and TOSA disallowing tensor with dimension
// 0. If not in place, something is very wrong.
@@ -190,65 +188,83 @@ TosaReduceTransposes::transposeDenseAttribute(DenseElementsAttr input,
return std::nullopt;
}
- if (input.isSplat())
+ auto newShape = applyTOSAPermutation(oldShape, perms);
+ RankedTensorType newType =
+ RankedTensorType::get(newShape, oldType.getElementType());
+
+ if (input.isSplat()) {
return input.reshape(newType);
+ }
+
+ auto rawData = input.getRawData();
+ if (!rawData.data()) {
+ return std::nullopt;
+ }
// The algorithm is approximately as follows:
- // input: perms, input flat array, input tensor type
- // (1/2) determine the strides of input/output if
- // they were strided in row-major order. (3) adjust the strides for the
- // input to be in the same order of indices as the output is written.
- // (4) process dimension by dimension. example: perms 2, 0, 1; input
- // 2x3x4; output 4x2x3 for i ... 4, j ... 2, k ... 3: output[i][j][k] =
- // input[j][k][i] output[6i + 3j + k] = input[12j + 4k + i] and we adjust
- // input strides to be as input[i + 12j + 4k] so we may process
- // layer-by-layer.
-
- // Step 1/2: Strides for input. We ignore output since row-major and can just
- // push_back.
-
- SmallVector<int64_t> originalInputStrides(rank);
- originalInputStrides[rank - 1] = 1;
- // index with int64_t to avoid overflow
- for (int64_t i = rank - 2; i >= 0; i--)
- originalInputStrides[i] =
- originalInputStrides[i + 1] * oldType.getDimSize(i + 1);
-
- // Step 3: Transpose strides of input to be same indexing (i, j, k, ...) as
- // output which is done in row-major order.
-
- SmallVector<int64_t> newInputStrides;
- newInputStrides.reserve(rank);
- for (int32_t v : perms)
- newInputStrides.push_back(originalInputStrides[v]);
-
- // Step 4: Write out the transposed "flat array" dimension by dimension.
-
- auto inputArray = input.getValues<Attribute>();
- SmallVector<std::pair<int64_t, int64_t>> boundsAndStrides;
- for (size_t i = 0; i < rank; i++)
- boundsAndStrides.push_back({newType.getDimSize(i), newInputStrides[i]});
-
- SmallVector<Attribute> resultArray;
- resultArray.reserve(inputArray.size());
-
- std::function<void(int64_t,
- SmallVector<std::pair<int64_t, int64_t>>::const_iterator)>
- processTransposeDim = [&](auto accumulatedIndex, auto it) {
- if (it == boundsAndStrides.end()) {
- resultArray.push_back(inputArray[accumulatedIndex]);
- return;
- }
-
- for (int64_t i = 0; i < it->first; i++) {
- int64_t j = accumulatedIndex + i * it->second;
- processTransposeDim(j, it + 1);
- }
- };
-
- processTransposeDim(0, boundsAndStrides.begin());
-
- return DenseElementsAttr::get(newType, resultArray);
+ // 1. Determine the strides of both input and output tensors in row-major
+ // order
+ // 2. Iterate through the output tensor linearly.
+ // 3. For each output position, decompose the linear index into
+ // multi-dimensional coordinates using output strides.
+ // 4. Use the permutation to map output coordinates to input coordinates and
+ // calculate the source linear index.
+
+ // Example: perms [2, 0, 1]; input 2x3x4; output 4x2x3
+ // for output linear index 11: decompose to output[1][1][2]
+ // using output strides [6,3,1]. Map to input coordinates using
+ // perms: dim 0→2, dim 1→0, dim 2→1, giving source position
+ // calculated as 1*inputStrides[2] + 1*inputStrides[0] + 2*inputStrides[1]
+ // = 1*1 + 1*12 + 2*4 = 21
+
+ size_t elementSize = oldType.getElementTypeBitWidth() / 8;
+ int64_t numElements = oldType.getNumElements();
+
+ SmallVector<char> outputBuffer(numElements * elementSize);
+ const char *inputPtr = rawData.data();
+ char *outputPtr = outputBuffer.data();
+
+ auto calculateStrides = [](ArrayRef<int64_t> shape) -> SmallVector<int64_t> {
+ int64_t rank = shape.size();
+ SmallVector<int64_t> strides(rank);
+ strides[rank - 1] = 1;
+ for (int64_t i = rank - 2; i >= 0; --i) {
+ strides[i] = strides[i + 1] * shape[i + 1];
+ }
+ return strides;
+ };
+
+ // Calculate strides for both input and output tensors
+ SmallVector<int64_t> inputStrides = calculateStrides(oldShape);
+ SmallVector<int64_t> outputStrides = calculateStrides(newShape);
+
+ auto mapCoordinates = [&](int64_t destLinearIndex) -> int64_t {
+ int64_t tempDestIndex = destLinearIndex;
+ int64_t sourceLinearIndex = 0;
+
+ // Decompose linear destination index into multi-dimensional
+ // coordinates dividing by output strides.
+ // Simultaneously map these coordinates through the permutation
+ // to calculate the corresponding source linear index.
+ for (auto j : llvm::seq<int64_t>(rank)) {
+ int64_t destCoord = tempDestIndex / outputStrides[j];
+ tempDestIndex %= outputStrides[j];
+ sourceLinearIndex += destCoord * inputStrides[perms[j]];
+ }
+
+ return sourceLinearIndex;
+ };
+
+ for (auto destLinearIndex : llvm::seq<int64_t>(numElements)) {
+ int64_t sourceLinearIndex = mapCoordinates(destLinearIndex);
+
+ // Copy the element from source to destination using type-agnostic byte
+ // copying.
+ std::memcpy(outputPtr + destLinearIndex * elementSize,
+ inputPtr + sourceLinearIndex * elementSize, elementSize);
+ }
+
+ return DenseElementsAttr::getFromRawBuffer(newType, outputBuffer);
}
// The SetVector should only contain ConstOp, ReshapeOp, TransposeOp
|
|
@sjarus , I couldn't find a good way to add a test for this change. Adding a performance test would be helpful, but I'm not sure about the best approach within the llvm-project. We could have used an end-to-end workflow testing approach if the changes were in |
There was a problem hiding this comment.
Thanks for this nice optimization!
Regarding testing, since this is a performance optimization, we couldn't accomplish anything further than the existing LIT tests in LLVM/MLIR.
Could you also enable this as default within TorchToTosa in Torch-MLIR, assuming that's your frontend ?
Hi Suraj, Thanks for the clarification regarding testing. Re: Could you also enable this as default within TorchToTosa in Torch-MLIR, assuming that's your frontend? This pass has already been added to the TorchToTosa backend as part of the PR below; I'm not sure if this is what you mean. Please let me know if you meant something else. |
|
@Hanumanth04 Congratulations on having your first Pull Request (PR) merged into the LLVM Project! Your changes will be combined with recent changes from other authors, then tested by our build bots. If there is a problem with a build, you may receive a report in an email or a comment on this PR. Please check whether problems have been caused by your change specifically, as the builds can include changes from many authors. It is not uncommon for your change to be included in a build that fails due to someone else's changes, or infrastructure issues. How to do this, and the rest of the post-merge process, is covered in detail here. If your change does cause a problem, it may be reverted, or you can revert it yourself. This is a normal part of LLVM development. You can fix your changes and open a new PR to merge them again. If you don't get any reports, no action is required from you. Your changes are working as expected, well done! |
The existing --tosa-reduce-transposes implementation is slow when transposing large constants (like weights in Linear operators) because it constructs expensive attributes during the transpose process.
For example, with a 1000×401408 weight tensor (network screen shot at the end for reference) , the current implementation takes:
This change addresses the performance issue by working directly with the raw tensor data, eliminating the need for costly intermediate attributes. With the changes made in this PR for the same tensor mentioned above, we take:
Network screen shot where this issue was observed: